(I'm mostly focused on Roman counts, but really information on any time period and any empire would be great.)
There are many important battles about which bias is clear, such as Thermopylae, which some sources claim involved over two million people. But with less notable battles, such as the dozens of clashes in Caesar's The Gallic War, are primary sources still reliable? And if we can't use them because they are too biased or just unreliable in some other way, how do we estimate?
I would approach the problem like this. At each instance ask yourself these questions:
How does the author know these casualty figures? Was he present at the battle, or did he have friends or relatives participate in it? Or perhaps is he describing something which happened three generations before him? How distant was he from the battlefield? If he was not present, where could he have gotten his numbers, and how accurate was that source?
What is the author’s purpose? Is he giving a history of Rome like Livy, or is he justifying Christianity like the Church Fathers, or is he perhaps writing propaganda like Caesar? Would it be to his advantage to inflate the figures, or perhaps to deflate them?
What was the author’s background? Would he know something about warfare because he was a former officer?
Does this statistic make common sense? Based upon what we know about the situation and the combatants, is it conceivable that this many soldiers were killed on this day?
How did this text get from the author to me? Was it translated from Latin to Greek to Arabic back to Latin and finally to English? Are there many lacunas? Could there have been transcription problems, like some medieval scribe accidently adding or subtracting from the figure? Is there great variability in the different versions of the text? Do scholars have problems with the text, or do they believe it is relatively intact?
Are there other primary sources which describe this battle? Do they tend to agree?
In reality, we probably can’t determine if a casualty figure was completely accurate or not. But we can make educated guesses.